I wrote: > Stuart Barkley <[email protected]> writes: > >> We also wanted this. Our primary use case was users needing to mmap() >> large files. > > What's the connexion? The PSS-based accounting was partly to address > that, though it can cause significant execd load for crazily fragmented > allocations (as seen particularly in proprietary FEM programs).
Sorry, I don't know what I was thinking of. Of course the rlimit counts named maps, not just anonymous ones (as you might hope it would), which is probably the main reason you might want to turn it off. The PSS limit prevents things being killed spuriously by execd monitoring due to buffers etc. accounted to the process. That was Mark's original complaint as far as I remember. [By the way, I think https://github.com/google/lmctfy/ had the right approach within the limits of cgroups and lacking a C interface. Unfortunately it seems dead in favour of spawn-of-Docker. It also won't build on RHEL6-ish systems, at least.] _______________________________________________ SGE-discuss mailing list [email protected] https://arc.liv.ac.uk/mailman/listinfo/sge-discuss
